CSDN

sklearn.ensemble._weight_boosting.BaseWeightBoosting.fit

for iboost in range(self.n_estimators):
    # Boosting step
    sample_weight, estimator_weight, estimator_error = self._boost(
        iboost,
        X, y,
        sample_weight,
        random_state)

    sample_weight_sum = np.sum(sample_weight)

    if iboost < self.n_estimators - 1:
        # Normalize
        sample_weight /= sample_weight_sum

sklearn.ensemble._weight_boosting.AdaBoostClassifier._boost_discrete

estimator.fit(X, y, sample_weight=sample_weight)

sample_weight就是$w_{mi}$，根据样本权重进行拟合

# Instances incorrectly classified
incorrect = y_predict != y

# Error fraction
estimator_error = np.mean(
    np.average(incorrect, weights=sample_weight, axis=0))

误分类率是误分类样本权重和： $e_m={\sum }_{i=1}^N{w}_{mi}I(G_m(x_i)\ne y_i)$

如果弱学习器的效果连随机都如不，早停。

计算$G_m$的系数${\alpha }_m=\frac{1}{2}log\frac{1-e_m}{e_m}$

SAMME算法考虑了多分类，乘以learning_rate的衰减实现正则化。

$w_{mi}=exp(-y_i{\alpha }_{m-1}{G}_{m-1}(x_i))$

试想，$y\times \hat{y}$，二分类时只有二者不同时才为1

这么写其实也兼顾了多分类的情况。对于多分类其实只需要修改${\alpha }_m$也就是代码中的estimator_weight的计算而已。